About the value of loss_w

[azuma164]

Dear author.
I found your work so interesting.
Then, I have a question about training settings to reproduce the results in your paper.
Can you share each value of loss_w related to pp, respectively?
I really appreciate any help you can provide.

[steb6]

Hi, did you solve this issue? I just tried with loss_w=0.03 with pp=0.5 on VOC2007 and I got about 93.2 (close enough to 93.6)

[jianbohuang]

hi, I run this code to reproduce the zero shot result in coco.
The F1-score of 46.35 at top_3 is significantly lower than the 50.3 reported in your paper.

train_zsl.py --config_file configs/models/rn50_ep50.yaml --datadir /PATH_COCO/ --dataset_config_file configs/datasets/coco.yaml --input_size 224 --lr 0.002 --loss_w 0.01 --n_ctx_pos 64 --n_ctx_neg 64

The best result for the unseen ZSL was:

Test: [5/50] p_unseen 32.50 r_unseen 80.76 f_unseen 46.35 mAP_unseen 64.03

Can you share your full implementation args for result in Table.4?